Research On Two Algorithms For Cost Sensitive Feature Selection

With the rapid development of internet technology,there are a lot of complex data in the practical application.It needs more money,time,resources and other costs to obtain useful data from these complex data.Therefore,in order to obtain useful data from these complex data,we should pay more money,time,resources and cost etc.Under the constraint cost how to obtain the most effective information from high dimensional data,multiple tags has become a hot issue in the research of data mining.In recent years,the test cost-sensitive learning plays a huge role in data mining.The test cost feature selection is a typical problem in the test cost-sensitive learning.Its aims is to minimize test cost,misclassification cost or total cost.In order to solve this problem,the scholars represent heuristic search algorithm,intelligent optimization algorithm and so on.However,these algorithms only converge to the local optimal or these algorithms efficiency is not satisfactory.Non-negative Matrix Factorization and multi-label learning play an important role in effectively dealing with the huge number of data.It becomes a new research direction in the field of high dimensional data analysis.Therefore,Based on the in-depth analysis and research of the non-negative matrix factorization algorithm,multi-label learning algorithms and cost sensitive learning,we combine the multi-label learning with cost-sensitive by using the non-negative matrix factorization in cost-sensitive feature selection algorithm.Non-negative matrix factorization cost-sensitive feature selection algorithm and multi-label under the background of cost sensitive feature selection algorithm are proposed.Two aspects are mainly studied in this paper.On the one hand,based on test cost-sensitive learning,we propose the cost-sensitive feature selection problem and find the test cost-sensitive feature selection method with the non-negative matrix factorization.First,the core technology is to generate a number of initial cost solutions using stochastic mechanisms,and form the a cost matrix.Then we define thequality of the fusion approximation based on matrix decomposition and design the fitness function of the test cost.Finally,we solve the problem by using the principle of iterative and select the best features subset.Experimental results show that the algorithm has a good performance in the large data set and has a low test cost.On the other hand,based on cost sensitive feature selection model in single label,we propose the transformation between the single label learning and multi-label learning.In addition,a multi-label test cost-sensitive feature selection algorithm is designed.This algorithm gives greater weight to the higher cost of samples to improve the higher cost for a class of sample prediction accuracy and choose the optimal feature subset.Therefore,the total cost will finally reduces.